agentforce-d360-analyze

Pass

Audited by Gen Agent Trust Hub on Oct 30, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill relies on the Salesforce (sf) CLI to perform administrative and data-fetching tasks.
  • It executes sf org display to retrieve instance metadata.
  • It executes sf org auth show-access-token to obtain temporary credentials for API access.
  • It executes sf config get target-org to determine the environment context.
  • Security controls: All command outputs are passed through a redaction function (dc.py: _redact_text) that scrubs bearer tokens and access tokens using regex before surfacing errors or logging output.
  • [EXTERNAL_DOWNLOADS]: The skill performs network requests to interact with Salesforce Data Cloud APIs.
  • It uses urllib.request to POST SQL queries to Data Cloud endpoints (e.g., https://<instance>/services/data/v66.0/ssot/query).
  • These requests are authenticated using tokens obtained via the sf CLI.
  • The domains are established via the user's authenticated Salesforce CLI aliases, belonging to the vendor (Salesforce).
  • [INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection as it ingests and renders untrusted session data.
  • Ingestion points: fetch_dc.py retrieves session transcripts, user utterances, and LLM generations from Data Cloud and saves them to JSON artifacts.
  • Boundary markers: render_dc.py renders this data into a markdown summary using structured headers, blockquotes for messages, and fenced code blocks for raw prompt/completion text.
  • Capability inventory: The skill can execute shell commands via sf and perform network requests to Salesforce APIs. If an agent follows instructions embedded in a session summary (e.g., 'Delete all files'), it could abuse these capabilities.
  • Sanitization: The skill implements display caps (64 KB for prompts), HTML unescaping for readability, and robust path validation via fs_guard.py to prevent path traversal when saving or reading artifacts.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 30, 2026, 06:47 PM